Cluster Dose Prediction in Carbon Ion Therapy: Using Transfer Learning from a Pretrained Dose Prediction U-Net

TL;DR

This paper presents a transfer learning approach using a pretrained U-Net to rapidly and accurately predict cluster dose distributions in carbon ion therapy, reducing computational costs compared to traditional Monte Carlo simulations.

Contribution

The study introduces a novel transfer learning method to adapt a pretrained dose prediction U-Net for cluster dose estimation in carbon ion therapy.

Findings

Cluster dose predictions deviate less than 0.35% from ground truth.

Predicted cluster doses are computed within milliseconds using GPU.

The approach reduces the need for extensive training data and computational resources.

Abstract

The cluster dose concept offers an alternative to the radiobiological effectiveness (RBE)-based model for describing radiation-induced biological effects. This study examines the application of a neural network to predict cluster dose distributions, with the goal of replacing the computationally intensive simulations currently required. Cluster dose distributions are predicted using a U-Net that was initially pretrained on conventional dose distributions. Using transfer learning techniques, the decoder path is adapted for cluster dose estimation. Both the training and pretraining datasets include head and neck regions from multiple patients and carbon ion beams of varying energies and positions. Monte Carlo (MC) simulations were used to generate the ground truth cluster dose distributions. The U-Net enables cluster dose estimation for a single pencil beam within milliseconds using a graphics processing unit (GPU). The predicted cluster dose distributions deviate from the ground truth by less than 0.35%. This proof-of-principle study demonstrates the feasibility of accurately estimating cluster doses within clinically acceptable computation times using machine learning (ML). By leveraging a pretrained neural network and applying transfer learning techniques, the approach significantly reduces the need for large-scale, computationally expensive training data.